from data centers to fog computing (part 1)...

Advanced Cloud Infrastructures

From Data Centers to Fog Computing (part 1)Guillaume Pierre

Master 2 CCS & SIF, 2017

Advanced Cloud Infrastructures 1 / 35

Table of Contents

1 Cloud Data Centers

2 Cloud-based applications are changing

3 Edge Computing

4 Fog Computing

5 Research question: what is the smallest, cheapest machine which canstill be used as a fog computing node?

6 Virtualizing a Raspberry Pi

Advanced Cloud Infrastructures Cloud Data Centers 4 / 35

Traditional system architectures

MachineOS

Middleware

Application

Traditionalarchitecture


The varying capacity problem

Com

pute

cap

aci

ty

Time

Demand



Com

pute

cap

aci

ty

Time

DemandNecessary capacity



Compute

capacity

Time




Compute

capacity

Time


What if demand increases beyond the capacity?


Cloud Computing

Di�cult to vary capacity!

Manual resourcemanagement

Resources available on demand

Resource management is fullyautomated

Pay only for what you use

VirtualizationMachine

OS

Middleware

Application

Machine + OSMachine + OSMachine + OS

Infrastructure as a Service


Cloudarchitecture


Cloud Computing

VirtualizationMachine

OS

Middleware

Application

Machine + OSMachine + OSMachine + OS

Infrastructure as a Service

Applications & Services


Cloudarchitecture

PaaS MapReduce Spark ...Storm


Table of Contents



3 Edge Computing

4 Fog Computing



Advanced Cloud Infrastructures Cloud-based applications are changing 8 / 35

The new mobile computing landscape

Since 2016: mobile network tra�c > �xed tra�c


New types of mobile applications

Interactive applications require ultra-low network latencies

E.g., augmented reality require end-to-end delays under 20ms

But latencies to the closest data center are 20-30ms using wirednetworks, up to 50-150ms on 4G mobile networks!!!


The new IoT landscape

Fact: IoT-generated tra�c grows faster than the Internet backbonecapacity


New types of mobile applications

Throughput-oriented applications require local computations

E.g., distributed videosurveillance is relevant only close to the cameras

Why waste long-distance network resources?


Question

Who owns computing resources located

closest to the mobile end usersand the IoT devices?

Answer: Mobile network operators


Table of Contents



3 Edge Computing

4 Fog Computing



Advanced Cloud Infrastructures Edge Computing 14 / 35

Mobile network operator's revenues

Where does mobile network operator's revenues come from?I 1990's: Voice (not any more)I 2000's: Text/SMS (not any more)I 2010's: Data (not for very long. . . )I What's next?

Services!

Let's steal part of the cloud computing business. . .

I No cloud data center can be closer to the users than us!


Mobile network operator's revenues

Where does mobile network operator's revenues come from?I 1990's: Voice (not any more)I 2000's: Text/SMS (not any more)I 2010's: Data (not for very long. . . )I What's next? Services!

Let's steal part of the cloud computing business. . .

I No cloud data center can be closer to the users than us!


Edge computing

Before:

After:


Standardization e�orts

European Telecommunications Standards Institute:Mobile Multi-Access Edge Computing

I Fujitsu, Hewlett-Packard, Huawei, Intel, Juniper, Motorola,NEC,Nokia, Orange, Samsung, Sony, Vodafone, . . .

I Focus: integration within mobile phone networks

Open Edge Computing InitiativeI Focus: opennessI Intel, Huawei, Vodafone, Carnegie-Mellon university, . . .

Open Fog ConsortiumI ARM, Cisco, Dell, Intel, Microsoft, . . .

Fog/Edge/Massively Distributed Clouds WG at OpenStack


Table of Contents



3 Edge Computing

4 Fog Computing



Advanced Cloud Infrastructures Fog Computing 18 / 35

What is fog computing?

Fog computing = cloud

+ edge

+ end-user devicesas a single execution platform

Low latency

Localized tra�cI Less global tra�cI Better reliabilityI PrivacyI . . .


Fog 6= Cloud

We need cloud servers close to the users, but the users are everywhere(and they are mobile)

I Let's place one cloud server within short range of any end-user device⇒ Fog computing resources will need to be distributed in thousands of

locations

Fog nodes will be connected with commodity networksI Forget your multi-Gbps data center networks!

Fog nodes must be small, cheap, easy to deploy and to replace


Fog 6= Cloud

Typical Cloud platform

Few data centers

Lots of resources per datacenter

High-performance networks

Cloud resource location is(mostly) irrelevant

Typical Fog platform

Huge number ofpoints-of-presence

Few resources per PoP

Commodity networks

Fog resource location isextremely important


Table of Contents



3 Edge Computing

4 Fog Computing



Advanced Cloud InfrastructuresResearch question: what is the smallest, cheapest machine

which can still be used as a fog computing node? 22 / 35

Raspberry PIs are more powerful than you may think

RPi3 Pine A64+ HP 820 G2

CPU (s) 46.5 4.1 2.8

Memory (MB/s) 933 1098 5658

Network (Mb/s) 94.2 922 935

Storage (MB/s) 2.53 2.42 23.9

Power when idle (W) 2 2 15

Power under load (W) 4.4 4.1 24.5

Price ∼ 92 e ∼ 74 e ∼ 1600 e

We chose RPi3 as our base platform for the time being

Mostly because of easy purchase options and community support. . .



What about a real application?

Input: live video stream

Processing: face recognition algorithm running inside Apache Flink

The RPIs are �only� 3-5 times slower than my laptop

But they are 17 times cheaper

If my applications scale horizontally I can use as many RPIs asnecessary



Table of Contents



3 Edge Computing

4 Fog Computing



Advanced Cloud Infrastructures Virtualizing a Raspberry Pi 25 / 35

Virtualization

Operating system

App 1 App 2 ...App 3

"The hardware machine"

Assembler layer

Digital logic layer (memory, compute circuits)

Micro-architecture

Peripherals (disk, screen,network, keyboard...)

Bus

Traditional machinearchitecture:

Applications

Operating system

Hardware


Virtualization

Operating system

Assembler layer


Assembler layer


Micro-architecture


Bus

Let's create a �special appli-cation� which behaves exactlythe same as the assemblerlayer. . .


Virtualization

Host operating system

Hypervisor(assembler layer)

Virtual machine

Assembler layer


Micro-architecture


Bus

Guest operating system

App 1 App 2 ...


We can execute any operatingsystem on top of it. . .

. . . and any applicationover the guest operatingsystem

⇒ We have a virtual ma-chine


Virtualization

Host operating system

Hypervisor

Virtual machine 1

Assembler layer

Digital logic layer(memory, compute circuits)

Micro-architecture


Bus


App 1 App 2 ...


Hypervisor

Virtual machine 2


App 1 App 2 ...

8 cores, 16 GB RAM

2 cores, 6 GB RAM 4 cores, 8 GB RAM

We can run multiple virtualmachines on the same physicalmachine:

Each virtual machineruns in full isolation fromthe other VMs

Each virtual machineowns a subset of thehardware resources of thephysical machine


Why is virtualization interesting for cloud providers?

Isolation: I can create multiple VMs on the same machine and giveeach VM to a di�erent user (they will not see nor interferewith each other)

Customization: Each user can customize their VMs according to their ownrequirements.

Consolidation: Few applications can really exploit a large server machine toits maximum capacity. With virtualization I can split thiscapacity in smaller units and thereby increase my resourceutilization.

Management: Virtualization simpli�es resource management: I canmeasure how many resources each user is using, migrate VMsfrom one host to another, etc.


Virtualization technologies

Virtualization technologies are now totally mainstream:

Commercial: VMware, Microsoft App-V, . . .

Open-Source: KVM, VirtualBox, Xen, . . .

Paravirtualization vs. full virtualization:

Paravirtualization works on any hardware platform but it requiresspecial support in the guest OS. Slow!

Full virtualization exploits special features of modern CPUs, does notrequire special support in the guest OS. Faster!

Virtual machines are waaaaaayyyyyy too heavyweight for a RPI

Each guest OS needs lots of memory

Each OS needs to execute lots of background stu�

Impossible to run 100+ VMs on a single machine. . .


Containers

Operating system



Assembler layer


Micro-architecture


Bus

Traditional machinearchitecture:

Applications

Operating system

Hardware


Containers

Operating system



Assembler layer


Micro-architecture


Bus

Container 1 Container 2

Let's create groups of processes whichbelong together:

Process groups are totally isolatedfrom each other

Each process group belongs to asingle user

Each process group has its ownhardware resource limits (CPU,RAM, . . . )

Each process group has its ownnetwork access policy

Etc.

⇒ We have containers


Container technologies

Containers are an operating system featureI No need for special CPU supportI Fully supported in Linux, Windows, . . .I We usually add an extra software layer to simplify management: Docker

Containers are less customizable than VMsI Container owners cannot choose their OSI But was that really necessary in the �rst place? Not always.

Containers are much more lightweight than VMsI No need to run lots of (mostly identical) operating systems next to

each otherI Containers often start in less than 1 secondI We can easily run hundreds of containers on a mid-sized machine


LXC containers

Performance

Performance is extremely closeto bare-metal

Agility

Starting 150 containers w/ ApacheI Total time: 36 s

(240ms/container)I Consumes about 2% of CPUI Memory usage: ∼ 10MB/container

Stopping 150 containers:I Total time: 9 seconds

http://www.slideshare.net/BodenRussell/realizing-linux-containerslxc


http://www.slideshare.net/BodenRussell/realizing-linux-containerslxc

Docker is a shipping container system for code

https://impythonist.wordpress.com/2015/06/21/

docker-the-future-of-virtualization-for-your-django-web-development/


https://impythonist.wordpress.com/2015/06/21/docker-the-future-of-virtualization-for-your-django-web-development/

https://impythonist.wordpress.com/2015/06/21/docker-the-future-of-virtualization-for-your-django-web-development/

https://blog.docker.com/2015/10/raspberry-pi-dockercon-challenge-winner/


https://blog.docker.com/2015/10/raspberry-pi-dockercon-challenge-winner/

Teaser

Next week we'll see thatbuilding a real fog platformis not as easy as it looks . . .


from data centers to fog computing (part 1)...

Documents